/*******************************************************
* Tune all the threshold of MPFR *
* Warning: tune the function in their dependent order!*
*******************************************************/
static void
all (void)
{
FILE *f = stdout;
time_t start_time, end_time;
struct tm *tp;
speed_time_init ();
if (verbose) {
printf ("Using: %s\n", speed_time_string);
printf ("speed_precision %d", speed_precision);
if (speed_unittime == 1.0)
printf (", speed_unittime 1 cycle");
else
printf (", speed_unittime %.2e secs", speed_unittime);
if (speed_cycletime == 1.0 || speed_cycletime == 0.0)
printf (", CPU freq unknown\n");
else
printf (", CPU freq %.2f MHz\n\n", 1e-6/speed_cycletime);
}
time (&start_time);
tp = localtime (&start_time);
fprintf (f, "/* Generated by MPFR's tuneup.c, %d-%02d-%02d, ",
tp->tm_year+1900, tp->tm_mon+1, tp->tm_mday);
#ifdef __ICC
fprintf (f, "icc %d.%d.%d */\n", __ICC / 100, __ICC / 10 % 10, __ICC % 10);
#elif defined(__GNUC__)
fprintf (f, "gcc %d.%d */\n", __GNUC__, __GNUC_MINOR__);
#elif defined (__SUNPRO_C)
fprintf (f, "Sun C %d.%d */\n", __SUNPRO_C / 0x100, __SUNPRO_C % 0x100);
#elif defined (__sgi) && defined (_COMPILER_VERSION)
fprintf (f, "MIPSpro C %d.%d.%d */\n",
_COMPILER_VERSION / 100,
_COMPILER_VERSION / 10 % 10,
_COMPILER_VERSION % 10);
#elif defined (__DECC) && defined (__DECC_VER)
fprintf (f, "DEC C %d */\n", __DECC_VER);
#else
fprintf (f, "system compiler */\n");
#endif
fprintf (f, "\n");
/* Tune mpfr_mul (threshold is in limbs, but it doesn't matter too much) */
if (verbose)
printf ("Measuring mpfr_mul with mpfr_mul_threshold=%lu...\n",
mpfr_mul_threshold);
tune_simple_func (&mpfr_mul_threshold, speed_mpfr_mul,
2*GMP_NUMB_BITS+1, 1000);
/* End of tuning */
time (&end_time);
if (verbose)
printf ("Complete (took %ld seconds).\n", end_time - start_time);
}
void start_timing(void)
{
if(seconds_per_cycle == 0.0)
speed_time_init();
if(need_cycles)
{
lock_thread_to_core();
start = __rdtsc();
}
else
{ LARGE_INTEGER ll;
QueryPerformanceCounter(&ll);
start = seconds_per_tick * ll.QuadPart;
}
}
int main(void)
{ char b1[_MAX_PATH], b2[_MAX_PATH];
unsigned long long cps;
if(get_processor_info(b1, b2, &cps) == EXIT_SUCCESS)
{
speed_time_init();
printf("\n%s", b1);
printf("\n%s", b2);
printf("\nSpeed: %lld", cps);
printf("\nSpeed: %.0f", 1.0 / seconds_per_cycle);
return 0;
}
else
return EXIT_FAILURE;
}
int
main (int argc, char *argv[])
{
int i;
int opt;
/* Unbuffered so output goes straight out when directed to a pipe or file
and isn't lost on killing the program half way. */
setbuf (stdout, NULL);
for (;;)
{
#if _GNU_SOURCE
opt = getopt(argc, argv, "a:CcDd::EFf:o:p:P:r::Rs:t:ux:y:w:W:z");
#else
opt = getopt(argc, argv, "a:CcDd:EFf:o:p:P:r:Rs:t:ux:y:w:W:z");
#endif
if (opt == EOF)
break;
switch (opt) {
case 'a':
if (strcmp (optarg, "random") == 0) option_data = DATA_RANDOM;
else if (strcmp (optarg, "random2") == 0) option_data = DATA_RANDOM2;
else if (strcmp (optarg, "zeros") == 0) option_data = DATA_ZEROS;
else if (strcmp (optarg, "aas") == 0) option_data = DATA_AAS;
else if (strcmp (optarg, "ffs") == 0) option_data = DATA_FFS;
else if (strcmp (optarg, "2fd") == 0) option_data = DATA_2FD;
else
{
fprintf (stderr, "unrecognised data option: %s\n", optarg);
exit (1);
}
break;
case 'C':
if (option_unit != UNIT_SECONDS) goto bad_unit;
option_unit = UNIT_CYCLESPERLIMB;
break;
case 'c':
if (option_unit != UNIT_SECONDS)
{
bad_unit:
fprintf (stderr, "cannot use more than one of -c, -C\n");
exit (1);
}
option_unit = UNIT_CYCLES;
break;
case 'D':
if (option_cmp != CMP_ABSOLUTE) goto bad_cmp;
option_cmp = CMP_DIFFPREV;
break;
case 'd':
if (option_cmp != CMP_ABSOLUTE)
{
bad_cmp:
fprintf (stderr, "cannot use more than one of -d, -D, -r\n");
exit (1);
}
option_cmp = CMP_DIFFERENCE;
option_cmp_pos=0;
if(optarg!=0)option_cmp_pos=atoi(optarg)-1;
break;
case 'E':
option_square = 1;
break;
case 'F':
option_square = 2;
break;
case 'f':
option_factor = atof (optarg);
if (option_factor <= 1.0)
{
fprintf (stderr, "-f factor must be > 1.0\n");
exit (1);
}
break;
case 'o':
speed_option_set (optarg);
break;
case 'P':
option_gnuplot = 1;
option_gnuplot_basename = optarg;
break;
case 'p':
speed_precision = atoi (optarg);
break;
case 'R':
option_seed = time (NULL);
break;
case 'r':
if (option_cmp != CMP_ABSOLUTE)
goto bad_cmp;
option_cmp = CMP_RATIO;
option_cmp_pos=0;
if(optarg!=0)option_cmp_pos = atoi(optarg)-1;
break;
case 's':
{
char *s;
for (s = strtok (optarg, ","); s != NULL; s = strtok (NULL, ","))
{
if (size_num == size_allocnum)
{
size_array = (struct size_array_t *)
__gmp_allocate_or_reallocate
(size_array,
size_allocnum * sizeof(size_array[0]),
(size_allocnum+10) * sizeof(size_array[0]));
size_allocnum += 10;
}
size_array[size_num].inc = 0;
if (sscanf (s, "%ld(%ld)%ld",
&size_array[size_num].start,
&size_array[size_num].inc,
&size_array[size_num].end) != 3)
{
if (sscanf (s, "%ld-%ld",
&size_array[size_num].start,
&size_array[size_num].end) != 2)
{
size_array[size_num].start = size_array[size_num].end
= atol (s);
}
}
if (size_array[size_num].start < 0
|| size_array[size_num].end < 0
|| size_array[size_num].start > size_array[size_num].end)
{
fprintf (stderr, "invalid size parameter: %s\n", s);
exit (1);
}
size_num++;
}
}
break;
case 't':
option_step = atol (optarg);
if (option_step < 1)
{
fprintf (stderr, "-t step must be >= 1\n");
exit (1);
}
break;
case 'u':
option_resource_usage = 1;
break;
case 'z':
sp.cache = 1;
break;
case 'x':
sp.align_xp = atol (optarg);
check_align_option ("-x", sp.align_xp);
break;
case 'y':
sp.align_yp = atol (optarg);
check_align_option ("-y", sp.align_yp);
break;
case 'w':
sp.align_wp = atol (optarg);
check_align_option ("-w", sp.align_wp);
break;
case 'W':
sp.align_wp2 = atol (optarg);
check_align_option ("-W", sp.align_wp2);
break;
case '?':
exit(1);
}
}
if (optind >= argc)
{
usage ();
exit (1);
}
if (size_num == 0)
{
fprintf (stderr, "-s <size> must be specified\n");
exit (1);
}
gmp_randinit_default (__gmp_rands);
__gmp_rands_initialized = 1;
gmp_randseed_ui (__gmp_rands, option_seed);
choice = (struct choice_t *) (*__gmp_allocate_func)
((argc - optind) * sizeof(choice[0]));
for ( ; optind < argc; optind++)
{
struct choice_t c;
routine_find (&c, argv[optind]);
choice[num_choices] = c;
num_choices++;
}
if ((option_cmp == CMP_RATIO || option_cmp == CMP_DIFFERENCE) &&
num_choices < 2)
{
fprintf (stderr, "WARNING, -d or -r does nothing when only one routine requested\n");
}
speed_time_init ();
if (option_unit == UNIT_CYCLES || option_unit == UNIT_CYCLESPERLIMB)
speed_cycletime_need_cycles ();
else
speed_cycletime_need_seconds ();
if (option_gnuplot)
{
run_gnuplot (argc, argv);
}
else
{
if (option_unit == UNIT_SECONDS)
printf ("overhead %.9f secs", speed_measure (speed_noop, NULL));
else
printf ("overhead %.2f cycles",
speed_measure (speed_noop, NULL) / speed_cycletime);
printf (", precision %d units of %.2e secs",
speed_precision, speed_unittime);
if (speed_cycletime == 1.0 || speed_cycletime == 0.0)
printf (", CPU freq unknown\n");
else
printf (", CPU freq %.2f MHz\n", 1e-6/speed_cycletime);
printf (" ");
for (i = 0; i < num_choices; i++)
printf (" %*s", COLUMN_WIDTH, choice[i].name);
printf ("\n");
run_all (stdout);
}
if (option_resource_usage)
{
#if HAVE_GETRUSAGE
{
/* This doesn't give data sizes on linux 2.0.x, only utime. */
struct rusage r;
if (getrusage (RUSAGE_SELF, &r) != 0)
perror ("getrusage");
else
printf ("getrusage(): utime %ld.%06ld data %ld stack %ld maxresident %ld\n",
r.ru_utime.tv_sec, r.ru_utime.tv_usec,
r.ru_idrss, r.ru_isrss, r.ru_ixrss);
}
#else
printf ("getrusage() not available\n");
#endif
/* Linux kernel. */
{
char buf[128];
sprintf (buf, "/proc/%d/status", getpid());
if (access (buf, R_OK) == 0)
{
sprintf (buf, "cat /proc/%d/status", getpid());
system (buf);
}
}
}
return 0;
}
void
usage (void)
{
int i;
speed_time_init ();
printf ("Usage: speed [-options] -s size <routine>...\n");
printf ("Measure the speed of some routines.\n");
printf ("Times are in seconds, accuracy is shown.\n");
printf ("\n");
printf (" -p num set precision as number of time units each routine must run\n");
printf (" -s x [,x]... sizes to measure: x = <size> | <start-end> | <start(step)end>\n");
printf (" single sizes or ranges, sep with comma or use multiple -s\n");
printf (" -t step step through sizes by given amount\n");
printf (" -f factor step through sizes by given factor (eg. 1.05)\n");
#if _GNU_SOURCE
printf (" -r [col] show times as ratios of the routine in column col(default column 1)\n");
printf (" -d [col] show times as difference from the routine in column col(default column 1)\n");
#else
printf (" -r col show times as ratios of the routine in column col\n");
printf (" -d col show times as difference from the routine in column col\n");
#endif
printf (" -D show times as difference from previous size shown\n");
printf (" -c show times in CPU cycles\n");
printf (" -C show times in cycles per limb\n");
printf (" -u print resource usage (memory) at end\n");
printf (" -P name output plot files \"name.gnuplot\" and \"name.data\"\n");
printf (" -a <type> use given data: random(default), random2, zeros, aas, ffs, 2fd\n");
printf (" -x, -y, -w, -W <align> specify data alignments, sources and dests\n");
printf (" -o addrs print addresses of data blocks\n");
printf (" colsum=A+B+...+Z Sums the columns A,B,.. upto a max of %d columns\n",SUMMAX);
printf (" colfile=col,filename Use the data in column col from file filename\n");
printf ("\n");
printf ("float*routine prefix can be used to multiply a column by a scale factor float.\n\n");
printf ("If both -t and -f are used, it means step by the factor or the step, whichever\n");
printf ("is greater.\n");
printf ("If both -C and -D are used, it means cycles per however many limbs between a\n");
printf ("size and the previous size.\n");
printf ("\n");
printf ("After running with -P, plots can be viewed with Gnuplot or Quickplot.\n");
printf ("\"gnuplot name.gnuplot\" (use \"set logscale xy; replot\" at the prompt for\n");
printf ("a log/log plot).\n");
printf ("\"quickplot -s name.data\" (has interactive zooming, and note -s is important\n");
printf ("when viewing more than one routine, it means same axis scales for all data).\n");
printf ("\n");
printf ("The available routines are as follows.\n");
printf ("\n");
for (i = 0; i < numberof (routine); i++)
{
if (routine[i].flag & FLAG_R)
printf ("\t%s.r\n", routine[i].name);
else if (routine[i].flag & FLAG_R_OPTIONAL)
printf ("\t%s (optional .r)\n", routine[i].name);
else
printf ("\t%s\n", routine[i].name);
}
printf ("\n");
printf ("Routines with a \".r\" need an extra parameter, for example mpn_lshift.6\n");
printf ("r should be in decimal, or use 0xN for hexadecimal.\n");
printf ("\n");
printf ("Special forms for r are \"<N>bits\" for a random N bit number, \"<N>ones\" for\n");
printf ("N one bits, or \"aas\" for 0xAA..AA.\n");
printf ("\n");
printf ("Times for sizes out of the range accepted by a routine are shown as 0.\n");
printf ("The fastest routine at each size is marked with a # (free form output only)\n");
printf ("but columns that are summed are excluded.\n");
printf ("\n");
printf ("%s", speed_time_string);
printf ("\n");
printf ("Gnuplot home page http://www.cs.dartmouth.edu/gnuplot_info.html\n");
printf ("Quickplot home page http://www.kachinatech.com/~quickplot\n");
}
void
usage (void)
{
int i;
speed_time_init ();
printf ("Usage: speed [-options] -s size <routine>...\n");
printf ("Measure the speed of some routines.\n");
printf ("Times are in seconds, accuracy is shown.\n");
printf ("\n");
printf (" -p num set precision as number of time units each routine must run\n");
printf (" -s size[-end][,size[-end]]... sizes to measure\n");
printf (" single sizes or ranges, sep with comma or use multiple -s\n");
printf (" -t step step through sizes by given amount\n");
printf (" -f factor step through sizes by given factor (eg. 1.05)\n");
printf (" -r show times as ratios of the first routine\n");
printf (" -d show times as difference from the first routine\n");
printf (" -D show times as difference from previous size shown\n");
printf (" -c show times in CPU cycles\n");
printf (" -C show times in cycles per limb\n");
printf (" -u print resource usage (memory) at end\n");
printf (" -P name output plot files \"name.gnuplot\" and \"name.data\"\n");
printf (" -a <type> use given data: random(default), random2, zeros, aas, ffs, 2fd\n");
printf (" -x, -y, -w, -W <align> specify data alignments, sources and dests\n");
printf (" -o addrs print addresses of data blocks\n");
printf ("\n");
printf ("If both -t and -f are used, it means step by the factor or the step, whichever\n");
printf ("is greater.\n");
printf ("If both -C and -D are used, it means cycles per however many limbs between a\n");
printf ("size and the previous size.\n");
printf ("\n");
printf ("After running with -P, plots can be viewed with Gnuplot or Quickplot.\n");
printf ("\"gnuplot name.gnuplot\" (use \"set logscale xy; replot\" at the prompt for\n");
printf ("a log/log plot).\n");
printf ("\"quickplot -s name.data\" (has interactive zooming, and note -s is important\n");
printf ("when viewing more than one routine, it means same axis scales for all data).\n");
printf ("\n");
printf ("The available routines are as follows.\n");
printf ("\n");
for (i = 0; i < numberof (routine); i++)
{
if (routine[i].flag & FLAG_R)
printf ("\t%s.r\n", routine[i].name);
else if (routine[i].flag & FLAG_R_OPTIONAL)
printf ("\t%s (optional .r)\n", routine[i].name);
else
printf ("\t%s\n", routine[i].name);
}
printf ("\n");
printf ("Routines with a \".r\" need an extra parameter, for example mpn_lshift.6\n");
printf ("r should be in decimal, or use 0xN for hexadecimal.\n");
printf ("\n");
printf ("Special forms for r are \"<N>bits\" for a random N bit number, \"<N>ones\" for\n");
printf ("N one bits, or \"aas\" for 0xAA..AA.\n");
printf ("\n");
printf ("Times for sizes out of the range accepted by a routine are shown as 0.\n");
printf ("The fastest routine at each size is marked with a # (free form output only).\n");
printf ("\n");
printf ("%s", speed_time_string);
printf ("\n");
printf ("Gnuplot home page http://www.gnuplot.info/\n");
printf ("Quickplot home page http://quickplot.sourceforge.net/\n");
}
